What Actually Is AI?

You probably used AI today without thinking about it.

We call all of these things artificial intelligence.

But they clearly aren't all doing the same thing.

So what actually makes something AI?

That's where we're starting.

With the most basic question underneath all of them:

What actually is artificial intelligence?

I know firsthand that it's possible to become a highly capable user of AI without fully understanding what's happening underneath.

Over the past few years, I've used AI to help me build businesses and websites, accelerate work that once took me days, learn unfamiliar subjects, and solve problems I wouldn't have immediately known how to approach.

And ironically, the more capable I became at using it, the more questions I had.

Every time I found myself wondering, Wait…can it do this too? — and discovering that the answer was often yes — I became more curious about what I was actually using.

What is AI? What is the difference between ChatGPT and the model powering it? Where does NVIDIA fit? Does an AI already know an answer, or does it search for one? How can a computer process language in the first place?

AI products are remarkably accessible. You don't have to understand how the technology works to benefit from it.

But I believe understanding how it works can help us use it better, recognize its limitations and risks, and imagine possibilities we might otherwise never think to explore.

That's why we're starting at the beginning.

The simplest useful definition

Artificial intelligence, or AI, is a broad field concerned with creating computer systems capable of performing tasks associated with intelligent behavior.

That can include things like:

• recognizing patterns
• understanding or generating language
• identifying objects in images
• making predictions
• recommending actions
• solving problems
• generating new content

Even defining AI precisely is surprisingly complicated. Different technical, research, and policy contexts draw its boundaries somewhat differently.

For our purposes, the most important thing to understand first is this:

ChatGPT is AI-powered.

But ChatGPT is not AI itself.

Generative AI is part of the modern AI landscape.

But generative AI is not all AI.

And machine learning is one enormously important approach to building AI systems.

But AI existed before modern machine learning became dominant.

Once you understand those distinctions, the rest of the AI world starts becoming much easier to map.

So what makes a computer system “intelligent”?

This is where the word intelligence can make AI sound more mysterious than it needs to.

Imagine that you want a computer system to look at thousands of transactions and identify which ones might be fraudulent.

A human might consider many things:

• Is this purchase unusually large?
• Is it happening somewhere unexpected?
• Does it resemble the person's normal spending?
• Does the combination of circumstances look suspicious?

We associate that kind of analysis with intelligence because it requires doing something more complex than simply storing a number.

An AI system might be designed to perform some version of that task.

That doesn't necessarily mean it thinks like the human analyst.

It means the system can perform a task that involves capabilities we associate with intelligent behavior.

And importantly, different AI systems can accomplish those tasks in very different ways.

AI existed before machines learned from data

Today, discussions about AI are dominated by machine learning.

But AI and machine learning are not synonyms.

For decades, researchers also built AI systems using approaches based on explicitly represented rules, logic, search, planning, and symbolic reasoning.

At a very simplified level, you can imagine one kind of rule-based system operating like this:

The rules can become extremely sophisticated, but humans are still explicitly defining much of the logic the system follows.

Then another approach became increasingly powerful:

Instead of trying to manually write every rule needed to solve a complicated problem, what if we could build systems that learn useful patterns from data?

That brings us to machine learning.

What does it mean for a machine to “learn”?

This is another phrase that can accidentally make computers sound magical.

When we say a machine-learning system learns, we don't necessarily mean it learns the way you do.

It doesn't have to sit there thinking:

Ah! Now I understand.

Instead, during a process called training, a machine-learning system is exposed to data and its internal numerical configuration is adjusted so that it becomes better at performing a particular task.

Very simplified:

That explanation raises about seventeen new questions. 😂

• How can words and pictures become something a computer can process?
• What exactly is a model?
• What is being “adjusted”?
• How does a computer find a pattern?

Excellent questions.

We're going to answer them.

But not all at once.

Because understanding AI gets much easier when we build the foundation in the right order.

Then where does deep learning fit?

Deep learning is a major approach within machine learning that uses neural networks with multiple computational layers.

Modern deep learning helped drive enormous advances in areas such as computer vision, speech recognition, language processing, and generative AI.

For now, you don't need to understand exactly how a neural network works.

You only need this mental model:

We'll eventually take apart what a neural network actually is — and, importantly, why the word neural does not mean we've simply built a digital human brain.

And then generative AI entered everyone's vocabulary

Many AI systems are designed primarily to do things like:

• classify — Is this email spam?
• predict — What product might this customer want?
• detect — Does this image contain a particular object?
• rank — Which search result is most relevant?
• recommend — What movie might this person enjoy?

Generative AI is designed to generate new outputs.

That might include:

• text
• images
• audio
• video
• software code

What is an LLM?

LLM stands for large language model.

An LLM is a type of AI model trained on large amounts of data to work with language and related patterns.

Models in this broad family can perform tasks such as generating text, summarizing information, answering questions, transforming writing, assisting with code, and much more.

GPT stands for Generative Pre-trained Transformer.

But here's one of the most important distinctions in this entire article:

Think of a model somewhat like an engine.

A company can build a larger product around that engine.

The product might add:

• a user interface
• memory
• search
• access to files
• external tools
• company data
• safety controls
• permissions
• specialized instructions

That larger system is what you actually interact with.

GPT is not ChatGPT

This distinction confused me when I first started trying to understand how all of this fit together — and I suspect I'm far from alone.

GPT refers to OpenAI's family of Generative Pre-trained Transformer models.

ChatGPT is the product you interact with, which uses underlying AI models along with additional product capabilities.

That's why the model powering a product can change while the product remains ChatGPT.

And the same broader principle applies elsewhere.

A product such as Microsoft Copilot can combine underlying models with Microsoft's interface, orchestration, tools, permissions, data connections, and product-specific capabilities.

Why does all of this terminology feel so confusing?

Partly because the AI world has developed an enormous vocabulary very quickly.

These terms are often used in the same conversation — and sometimes used interchangeably even when they describe completely different things.

Add in the fact that companies sometimes use the same brand name for both a model family and a product, and it's no wonder people struggle to build a clear mental map.

That's one reason AI, Actually. will keep asking a very simple question:

Wait — what category is that?

Because knowing the name of something is not the same as understanding what it is.

Imagine someone says:

“We're using Copilot with an OpenAI model running on NVIDIA hardware.”

Those names might all appear in one sentence.

But they aren't three competing versions of the same thing.

• Copilot → product
• OpenAI model → model
• NVIDIA hardware → compute/infrastructure

Understanding the category tells you what role each piece is playing.

And one correction to a question I had when I first began mapping this myself:

GPT is not an NVIDIA model.

OpenAI develops GPT models.

NVIDIA builds important computing technologies used across the AI industry.

Hearing two names constantly mentioned in AI conversations doesn't mean they belong to the same category.

So does ChatGPT “go to the internet” every time you ask something?

Not necessarily.

And this is another distinction we'll explore much more deeply later.

A language model can generate a response based on patterns learned during training and information available in the context it receives.

A product built around a model can also be given tools.

One of those tools might be web search.

Another might retrieve a document.

Another might run code.

Another might interact with another software system.

So:

• The model generating language
• and
• the product retrieving outside information
• can be different parts of the overall system.

This is why two products using related underlying models can still behave very differently.

The model matters.

But so do the systems built around it.

Is AI actually intelligent?

Now we arrive at one of the biggest questions in the entire field.

And the most accurate beginner answer is:

It depends on exactly what we mean by intelligence.

AI systems can demonstrate extraordinary capabilities.

They can identify patterns humans might miss.

They can generate useful language.

They can outperform people at particular tasks.

They can also make bizarre mistakes.

They can confidently produce false information.

They can fail when circumstances differ from what their design or training prepared them for.

And observing intelligent-seeming behavior does not, by itself, answer philosophical questions about consciousness, subjective experience, or whether a machine understands something in exactly the way a human does.

Those are different questions.

For now, the important distinction is:

And when we eventually reach the debate over whether language models are “just autocomplete” or something much more, we're going to do something slightly unusual:

We're going to understand how they actually work before deciding what we think about the hot take.

What Does This Mean for Me?

If you're a professional or business owner, there's a good chance you're already using AI — even if you think your organization hasn't “adopted AI” yet.

Your email system filtering spam?

AI in action.

A financial institution detecting potentially suspicious transactions?

AI in action.

A recommendation system deciding what content or product to show someone next?

Also AI.

Using a product like Microsoft Copilot, ChatGPT, or Claude to help with work?

Now you're interacting with another part of the AI landscape: an AI-powered product built around underlying models and other systems.

Understanding these distinctions matters because:

Before choosing a product, an organization should understand the problem it's trying to solve.

• What is the desired outcome?
• What happens today?
• Where is the actual pain point?
• What data is involved?
• What people, processes, and systems does the solution need to work with?
• What risks need to be governed?
• And how will we know whether the solution actually worked?

Only then can we intelligently ask what kind of technology — including what kind of AI — belongs in the solution.

Understanding what AI actually is gives you the foundation to start asking those better questions.

Why understand what's underneath?

You don't need to know how an internal combustion engine works to drive a car.

And you don't need to understand neural networks, transformers, or tokens to open an AI product and get something useful from it.

That's part of what makes this technology so exciting.

But understanding what's underneath changes what you're able to do with it.

You can ask better questions.

You can recognize more possibilities.

You can understand limitations and risks more clearly.

You can make better decisions about which tools belong where.

And when something new arrives — as it inevitably will — you have a foundation to connect it to instead of starting over every time.

I've always believed in learning from the basement level up. Not because everyone needs to become an engineer, but because the stronger your mental model of a system is, the more intentionally you can work with it.

But here's the thing:

Most of us didn't actually enter AI through the basement.

We walked right in through the front door.

Products like ChatGPT made incredibly sophisticated AI accessible to almost anyone with a text box and a question.

You could start using the technology before you understood almost anything about what was happening underneath.

And that's a wonderful thing.

But if you're curious about what makes the house stand — what's beneath the floorboards, how the rooms connect, what's powering everything behind the walls — sometimes it's worth taking a trip down to the basement.

Not because you need to live there. Because understanding the foundation helps you understand everything built on top of it.

That's what Learn AI with Kelly is for.

One layer at a time.

And if you just read that last sentence and thought:

Okay…but how does a computer do any of this?

Good.

Because we're going there.

But first, we need to understand how we got here.